1 //===-- PPCTargetMachine.cpp - Define TargetMachine for PowerPC -----------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // Top-level implementation for the PowerPC target. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "MCTargetDesc/PPCMCTargetDesc.h" 15 #include "PPC.h" 16 #include "PPCSubtarget.h" 17 #include "PPCTargetObjectFile.h" 18 #include "PPCTargetMachine.h" 19 #include "PPCTargetTransformInfo.h" 20 #include "llvm/ADT/Optional.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/ADT/Triple.h" 24 #include "llvm/Analysis/TargetTransformInfo.h" 25 #include "llvm/CodeGen/Passes.h" 26 #include "llvm/CodeGen/TargetPassConfig.h" 27 #include "llvm/IR/Attributes.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/Function.h" 30 #include "llvm/Pass.h" 31 #include "llvm/Support/CodeGen.h" 32 #include "llvm/Support/CommandLine.h" 33 #include "llvm/Support/TargetRegistry.h" 34 #include "llvm/Target/TargetLoweringObjectFile.h" 35 #include "llvm/Target/TargetOptions.h" 36 #include "llvm/Transforms/Scalar.h" 37 #include <cassert> 38 #include <memory> 39 #include <string> 40 41 using namespace llvm; 42 43 static cl:: 44 opt<bool> DisableCTRLoops("disable-ppc-ctrloops", cl::Hidden, 45 cl::desc("Disable CTR loops for PPC")); 46 47 static cl:: 48 opt<bool> DisablePreIncPrep("disable-ppc-preinc-prep", cl::Hidden, 49 cl::desc("Disable PPC loop preinc prep")); 50 51 static cl::opt<bool> 52 VSXFMAMutateEarly("schedule-ppc-vsx-fma-mutation-early", 53 cl::Hidden, cl::desc("Schedule VSX FMA instruction mutation early")); 54 55 static cl:: 56 opt<bool> DisableVSXSwapRemoval("disable-ppc-vsx-swap-removal", cl::Hidden, 57 cl::desc("Disable VSX Swap Removal for PPC")); 58 59 static cl:: 60 opt<bool> DisableQPXLoadSplat("disable-ppc-qpx-load-splat", cl::Hidden, 61 cl::desc("Disable QPX load splat simplification")); 62 63 static cl:: 64 opt<bool> DisableMIPeephole("disable-ppc-peephole", cl::Hidden, 65 cl::desc("Disable machine peepholes for PPC")); 66 67 static cl::opt<bool> 68 EnableGEPOpt("ppc-gep-opt", cl::Hidden, 69 cl::desc("Enable optimizations on complex GEPs"), 70 cl::init(true)); 71 72 static cl::opt<bool> 73 EnablePrefetch("enable-ppc-prefetching", 74 cl::desc("disable software prefetching on PPC"), 75 cl::init(false), cl::Hidden); 76 77 static cl::opt<bool> 78 EnableExtraTOCRegDeps("enable-ppc-extra-toc-reg-deps", 79 cl::desc("Add extra TOC register dependencies"), 80 cl::init(true), cl::Hidden); 81 82 static cl::opt<bool> 83 EnableMachineCombinerPass("ppc-machine-combiner", 84 cl::desc("Enable the machine combiner pass"), 85 cl::init(true), cl::Hidden); 86 87 extern "C" void LLVMInitializePowerPCTarget() { 88 // Register the targets 89 RegisterTargetMachine<PPC32TargetMachine> A(getThePPC32Target()); 90 RegisterTargetMachine<PPC64TargetMachine> B(getThePPC64Target()); 91 RegisterTargetMachine<PPC64TargetMachine> C(getThePPC64LETarget()); 92 93 PassRegistry &PR = *PassRegistry::getPassRegistry(); 94 initializePPCBoolRetToIntPass(PR); 95 initializePPCExpandISELPass(PR); 96 } 97 98 /// Return the datalayout string of a subtarget. 99 static std::string getDataLayoutString(const Triple &T) { 100 bool is64Bit = T.getArch() == Triple::ppc64 || T.getArch() == Triple::ppc64le; 101 std::string Ret; 102 103 // Most PPC* platforms are big endian, PPC64LE is little endian. 104 if (T.getArch() == Triple::ppc64le) 105 Ret = "e"; 106 else 107 Ret = "E"; 108 109 Ret += DataLayout::getManglingComponent(T); 110 111 // PPC32 has 32 bit pointers. The PS3 (OS Lv2) is a PPC64 machine with 32 bit 112 // pointers. 113 if (!is64Bit || T.getOS() == Triple::Lv2) 114 Ret += "-p:32:32"; 115 116 // Note, the alignment values for f64 and i64 on ppc64 in Darwin 117 // documentation are wrong; these are correct (i.e. "what gcc does"). 118 if (is64Bit || !T.isOSDarwin()) 119 Ret += "-i64:64"; 120 else 121 Ret += "-f64:32:64"; 122 123 // PPC64 has 32 and 64 bit registers, PPC32 has only 32 bit ones. 124 if (is64Bit) 125 Ret += "-n32:64"; 126 else 127 Ret += "-n32"; 128 129 return Ret; 130 } 131 132 static std::string computeFSAdditions(StringRef FS, CodeGenOpt::Level OL, 133 const Triple &TT) { 134 std::string FullFS = FS; 135 136 // Make sure 64-bit features are available when CPUname is generic 137 if (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le) { 138 if (!FullFS.empty()) 139 FullFS = "+64bit," + FullFS; 140 else 141 FullFS = "+64bit"; 142 } 143 144 if (OL >= CodeGenOpt::Default) { 145 if (!FullFS.empty()) 146 FullFS = "+crbits," + FullFS; 147 else 148 FullFS = "+crbits"; 149 } 150 151 if (OL != CodeGenOpt::None) { 152 if (!FullFS.empty()) 153 FullFS = "+invariant-function-descriptors," + FullFS; 154 else 155 FullFS = "+invariant-function-descriptors"; 156 } 157 158 return FullFS; 159 } 160 161 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 162 // If it isn't a Mach-O file then it's going to be a linux ELF 163 // object file. 164 if (TT.isOSDarwin()) 165 return llvm::make_unique<TargetLoweringObjectFileMachO>(); 166 167 return llvm::make_unique<PPC64LinuxTargetObjectFile>(); 168 } 169 170 static PPCTargetMachine::PPCABI computeTargetABI(const Triple &TT, 171 const TargetOptions &Options) { 172 if (Options.MCOptions.getABIName().startswith("elfv1")) 173 return PPCTargetMachine::PPC_ABI_ELFv1; 174 else if (Options.MCOptions.getABIName().startswith("elfv2")) 175 return PPCTargetMachine::PPC_ABI_ELFv2; 176 177 assert(Options.MCOptions.getABIName().empty() && 178 "Unknown target-abi option!"); 179 180 if (!TT.isMacOSX()) { 181 switch (TT.getArch()) { 182 case Triple::ppc64le: 183 return PPCTargetMachine::PPC_ABI_ELFv2; 184 case Triple::ppc64: 185 return PPCTargetMachine::PPC_ABI_ELFv1; 186 default: 187 // Fallthrough. 188 ; 189 } 190 } 191 return PPCTargetMachine::PPC_ABI_UNKNOWN; 192 } 193 194 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 195 Optional<Reloc::Model> RM) { 196 if (!RM.hasValue()) { 197 if (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le) { 198 if (!TT.isOSBinFormatMachO() && !TT.isMacOSX()) 199 return Reloc::PIC_; 200 } 201 if (TT.isOSDarwin()) 202 return Reloc::DynamicNoPIC; 203 return Reloc::Static; 204 } 205 return *RM; 206 } 207 208 // The FeatureString here is a little subtle. We are modifying the feature 209 // string with what are (currently) non-function specific overrides as it goes 210 // into the LLVMTargetMachine constructor and then using the stored value in the 211 // Subtarget constructor below it. 212 PPCTargetMachine::PPCTargetMachine(const Target &T, const Triple &TT, 213 StringRef CPU, StringRef FS, 214 const TargetOptions &Options, 215 Optional<Reloc::Model> RM, 216 CodeModel::Model CM, CodeGenOpt::Level OL) 217 : LLVMTargetMachine(T, getDataLayoutString(TT), TT, CPU, 218 computeFSAdditions(FS, OL, TT), Options, 219 getEffectiveRelocModel(TT, RM), CM, OL), 220 TLOF(createTLOF(getTargetTriple())), 221 TargetABI(computeTargetABI(TT, Options)), 222 Subtarget(TargetTriple, CPU, computeFSAdditions(FS, OL, TT), *this) { 223 224 initAsmInfo(); 225 } 226 227 PPCTargetMachine::~PPCTargetMachine() = default; 228 229 void PPC32TargetMachine::anchor() {} 230 231 PPC32TargetMachine::PPC32TargetMachine(const Target &T, const Triple &TT, 232 StringRef CPU, StringRef FS, 233 const TargetOptions &Options, 234 Optional<Reloc::Model> RM, 235 CodeModel::Model CM, 236 CodeGenOpt::Level OL) 237 : PPCTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} 238 239 void PPC64TargetMachine::anchor() {} 240 241 PPC64TargetMachine::PPC64TargetMachine(const Target &T, const Triple &TT, 242 StringRef CPU, StringRef FS, 243 const TargetOptions &Options, 244 Optional<Reloc::Model> RM, 245 CodeModel::Model CM, 246 CodeGenOpt::Level OL) 247 : PPCTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} 248 249 const PPCSubtarget * 250 PPCTargetMachine::getSubtargetImpl(const Function &F) const { 251 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 252 Attribute FSAttr = F.getFnAttribute("target-features"); 253 254 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 255 ? CPUAttr.getValueAsString().str() 256 : TargetCPU; 257 std::string FS = !FSAttr.hasAttribute(Attribute::None) 258 ? FSAttr.getValueAsString().str() 259 : TargetFS; 260 261 // FIXME: This is related to the code below to reset the target options, 262 // we need to know whether or not the soft float flag is set on the 263 // function before we can generate a subtarget. We also need to use 264 // it as a key for the subtarget since that can be the only difference 265 // between two functions. 266 bool SoftFloat = 267 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 268 // If the soft float attribute is set on the function turn on the soft float 269 // subtarget feature. 270 if (SoftFloat) 271 FS += FS.empty() ? "-hard-float" : ",-hard-float"; 272 273 auto &I = SubtargetMap[CPU + FS]; 274 if (!I) { 275 // This needs to be done before we create a new subtarget since any 276 // creation will depend on the TM and the code generation flags on the 277 // function that reside in TargetOptions. 278 resetTargetOptions(F); 279 I = llvm::make_unique<PPCSubtarget>( 280 TargetTriple, CPU, 281 // FIXME: It would be good to have the subtarget additions here 282 // not necessary. Anything that turns them on/off (overrides) ends 283 // up being put at the end of the feature string, but the defaults 284 // shouldn't require adding them. Fixing this means pulling Feature64Bit 285 // out of most of the target cpus in the .td file and making it set only 286 // as part of initialization via the TargetTriple. 287 computeFSAdditions(FS, getOptLevel(), getTargetTriple()), *this); 288 } 289 return I.get(); 290 } 291 292 //===----------------------------------------------------------------------===// 293 // Pass Pipeline Configuration 294 //===----------------------------------------------------------------------===// 295 296 namespace { 297 298 /// PPC Code Generator Pass Configuration Options. 299 class PPCPassConfig : public TargetPassConfig { 300 public: 301 PPCPassConfig(PPCTargetMachine *TM, PassManagerBase &PM) 302 : TargetPassConfig(TM, PM) {} 303 304 PPCTargetMachine &getPPCTargetMachine() const { 305 return getTM<PPCTargetMachine>(); 306 } 307 308 void addIRPasses() override; 309 bool addPreISel() override; 310 bool addILPOpts() override; 311 bool addInstSelector() override; 312 void addMachineSSAOptimization() override; 313 void addPreRegAlloc() override; 314 void addPreSched2() override; 315 void addPreEmitPass() override; 316 }; 317 318 } // end anonymous namespace 319 320 TargetPassConfig *PPCTargetMachine::createPassConfig(PassManagerBase &PM) { 321 return new PPCPassConfig(this, PM); 322 } 323 324 void PPCPassConfig::addIRPasses() { 325 if (TM->getOptLevel() != CodeGenOpt::None) 326 addPass(createPPCBoolRetToIntPass()); 327 addPass(createAtomicExpandPass(&getPPCTargetMachine())); 328 329 // For the BG/Q (or if explicitly requested), add explicit data prefetch 330 // intrinsics. 331 bool UsePrefetching = TM->getTargetTriple().getVendor() == Triple::BGQ && 332 getOptLevel() != CodeGenOpt::None; 333 if (EnablePrefetch.getNumOccurrences() > 0) 334 UsePrefetching = EnablePrefetch; 335 if (UsePrefetching) 336 addPass(createLoopDataPrefetchPass()); 337 338 if (TM->getOptLevel() >= CodeGenOpt::Default && EnableGEPOpt) { 339 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 340 // and lower a GEP with multiple indices to either arithmetic operations or 341 // multiple GEPs with single index. 342 addPass(createSeparateConstOffsetFromGEPPass(TM, true)); 343 // Call EarlyCSE pass to find and remove subexpressions in the lowered 344 // result. 345 addPass(createEarlyCSEPass()); 346 // Do loop invariant code motion in case part of the lowered result is 347 // invariant. 348 addPass(createLICMPass()); 349 } 350 351 TargetPassConfig::addIRPasses(); 352 } 353 354 bool PPCPassConfig::addPreISel() { 355 if (!DisablePreIncPrep && getOptLevel() != CodeGenOpt::None) 356 addPass(createPPCLoopPreIncPrepPass(getPPCTargetMachine())); 357 358 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None) 359 addPass(createPPCCTRLoops(getPPCTargetMachine())); 360 361 return false; 362 } 363 364 bool PPCPassConfig::addILPOpts() { 365 addPass(&EarlyIfConverterID); 366 367 if (EnableMachineCombinerPass) 368 addPass(&MachineCombinerID); 369 370 return true; 371 } 372 373 bool PPCPassConfig::addInstSelector() { 374 // Install an instruction selector. 375 addPass(createPPCISelDag(getPPCTargetMachine())); 376 377 #ifndef NDEBUG 378 if (!DisableCTRLoops && getOptLevel() != CodeGenOpt::None) 379 addPass(createPPCCTRLoopsVerify()); 380 #endif 381 382 addPass(createPPCVSXCopyPass()); 383 return false; 384 } 385 386 void PPCPassConfig::addMachineSSAOptimization() { 387 TargetPassConfig::addMachineSSAOptimization(); 388 // For little endian, remove where possible the vector swap instructions 389 // introduced at code generation to normalize vector element order. 390 if (TM->getTargetTriple().getArch() == Triple::ppc64le && 391 !DisableVSXSwapRemoval) 392 addPass(createPPCVSXSwapRemovalPass()); 393 // Target-specific peephole cleanups performed after instruction 394 // selection. 395 if (!DisableMIPeephole) { 396 addPass(createPPCMIPeepholePass()); 397 addPass(&DeadMachineInstructionElimID); 398 } 399 } 400 401 void PPCPassConfig::addPreRegAlloc() { 402 if (getOptLevel() != CodeGenOpt::None) { 403 initializePPCVSXFMAMutatePass(*PassRegistry::getPassRegistry()); 404 insertPass(VSXFMAMutateEarly ? &RegisterCoalescerID : &MachineSchedulerID, 405 &PPCVSXFMAMutateID); 406 } 407 408 // FIXME: We probably don't need to run these for -fPIE. 409 if (getPPCTargetMachine().isPositionIndependent()) { 410 // FIXME: LiveVariables should not be necessary here! 411 // PPCTLSDYnamicCallPass uses LiveIntervals which previously dependet on 412 // LiveVariables. This (unnecessary) dependency has been removed now, 413 // however a stage-2 clang build fails without LiveVariables computed here. 414 addPass(&LiveVariablesID, false); 415 addPass(createPPCTLSDynamicCallPass()); 416 } 417 if (EnableExtraTOCRegDeps) 418 addPass(createPPCTOCRegDepsPass()); 419 } 420 421 void PPCPassConfig::addPreSched2() { 422 if (getOptLevel() != CodeGenOpt::None) { 423 addPass(&IfConverterID); 424 425 // This optimization must happen after anything that might do store-to-load 426 // forwarding. Here we're after RA (and, thus, when spills are inserted) 427 // but before post-RA scheduling. 428 if (!DisableQPXLoadSplat) 429 addPass(createPPCQPXLoadSplatPass()); 430 } 431 } 432 433 void PPCPassConfig::addPreEmitPass() { 434 addPass(createPPCExpandISELPass()); 435 436 if (getOptLevel() != CodeGenOpt::None) 437 addPass(createPPCEarlyReturnPass(), false); 438 // Must run branch selection immediately preceding the asm printer. 439 addPass(createPPCBranchSelectionPass(), false); 440 } 441 442 TargetIRAnalysis PPCTargetMachine::getTargetIRAnalysis() { 443 return TargetIRAnalysis([this](const Function &F) { 444 return TargetTransformInfo(PPCTTIImpl(this, F)); 445 }); 446 } 447